Part 2 — Session guide
Friday, 23 October 2026. We use the 270-patient heart-disease dataset matching Wright et al. (2026), with 13 predictors and an absent/present outcome.
- Open
part2.Rproj; keeppart2/as the working directory. - Run
source("install-packages.R")once, thensource("check-setup.R"). - Follow ml-workflow-2.qmd.
- Run the scripts in
R/in article order.
| Script | Article workflow |
|---|---|
| 01-data.R | Frozen OpenML data and category definitions |
| 02-trees-and-evaluation.R | Trees, pruning, random forest, evaluation and resampling |
| 03-neural-network.R | Published neural-network architecture and training settings |
| 04-tuning.R | Grid/random/MBO tuning and nested cross-validation |
| 05-synthetic-data.R | ARF, FORDE and FORGE |
Render the core notebook with quarto render. Neural-network, full tuning and synthetic-data scripts are separate runs. Full tuning retains the article budgets and can take much longer than the two-hour session.
Keras setup
Install the R dependencies, then configure a supported Python backend using keras3::install_keras() in a fresh R session. Alternatively, select an existing compatible Python environment before loading Keras. See the official Keras for R setup guide. Backend installation is a separate step from installing the R package.
The core workflow, neural network, full random/MBO searches and synthetic-data example have been run. See validation results for the results and remaining full-budget checks.
Scope and provenance
The companion repository cited by the article currently returns 404. These scripts implement the published examples with explicit setup choices; they are not a verified copy of that repository. The data guide and replication record document what is specified, what was reconstructed and what has been tested. The article’s printed scores are not guaranteed for a different split or environment.